US12147577B2ActiveUtilityA1
Interpretability framework for differentially private deep learning
Est. expiryOct 30, 2040(~14.3 yrs left)· nominal 20-yr term from priority
G06N 3/09G06N 3/0464G06F 18/2148G06F 17/18G06N 20/00G06F 18/2413G06N 3/08G06F 21/6254G06F 21/6245
78
PatentIndex Score
0
Cited by
34
References
20
Claims
Abstract
Data is received that specifies a bound for an adversarial posterior belief ρ c that corresponds to a likelihood to re-identify data points from the dataset based on a differentially private function output. Privacy parameters ε, δ are then calculated based on the received data that govern a differential privacy (DP) algorithm to be applied to a function to be evaluated over a dataset. The calculating is based on a ratio of probabilities distributions of different observations, which are bound by the posterior belief ρ c as applied to a dataset. The calculated privacy parameters are then used to apply the DP algorithm to the function over the dataset. Related apparatus, systems, techniques and articles are also described.
Claims
exact text as granted — not AI-modifiedWhat is claimed is:
1. A computer-implemented method for anonymized analysis of datasets comprising:
receiving data specifying privacy parameters ε, δ which govern a differential privacy (DP) algorithm to be applied to a function to be evaluated over a dataset;
calculating, based on the received data, an expected membership advantage ρ α that corresponds to a likelihood of an adversary successfully identifying a member in the dataset, the calculating being based on an overlap of two probability distributions; and
applying, using the calculated expected membership advantage ρ α , the DP algorithm to a function over the dataset.
2. The method of claim 1 , wherein the probability distributions are generated using a Gaussian mechanism with an (ε, δ) guarantee that perturbs a result of the function evaluated over the dataset, ensuring that membership advantage is ρ α on the dataset.
3. The method of claim 1 , further comprising:
anonymously training at least one machine learning model using the dataset after application of the DP algorithm to the function over the dataset.
4. The method of claim 3 , further comprising:
deploying the trained at least one machine learning model to classify further data input into the at least one machine learning model.
5. The method of claim 1 , wherein the calculated expected membership advantage ρ α for a series of (ε, δ) anonymized function evaluations with multidimensional data is equal to:
CDF
(
1
2
2
ln
(
1.25
δ
)
ϵ
)
-
CDF
(
-
1
2
2
ln
(
1.25
δ
)
ϵ
)
wherein CDF is a cumulative distribution function of a standard normal distribution.
6. The method of claim 4 , further comprising:
calculating a resulting expected membership advantage ρ α using sequential composition or Rényi differential privacy (RDP) composition; and
updating the at least one machine learning model using the calculated resulting expected membership advantage ρ α .
7. The method of claim 1 , wherein the calculating is based on a conditional probability of different possible datasets.
8. A system comprising:
at least one hardware processor; and
memory storing an application executable by the at least one hardware processor of the system to perform operations comprising:
receiving data specifying privacy parameters ε, δ which govern a differential privacy (DP) algorithm to be applied to a function to be evaluated over a dataset;
calculating, based on the received data, an expected membership advantage ρ α that corresponds to a likelihood of an adversary successfully identifying a member in the dataset, the calculating being based on an overlap of two probability distributions; and
applying, using the calculated expected membership advantage ρ α , the DP algorithm to a function over the dataset.
9. The system of claim 8 , wherein the probability distributions are generated using a Gaussian mechanism with an (ε, δ) guarantee that perturbs a result of the function evaluated over the dataset, ensuring that membership advantage is ρ α on the dataset.
10. The system of claim 8 , wherein the operations further comprise:
anonymously training at least one machine learning model using the dataset after application of the DP algorithm to the function over the dataset.
11. The system of claim 10 , wherein the operations further comprise:
deploying the trained at least one machine learning model to classify further data input into the at least one machine learning model.
12. The system of claim 8 , wherein the calculated expected membership advantage ρ α for a series of (ε, δ) anonymized function evaluations with multidimensional data is equal to:
CDF
(
1
2
2
ln
(
1.25
δ
)
ϵ
)
-
CDF
(
-
1
2
2
ln
(
1.25
δ
)
ϵ
)
wherein CDF is a cumulative distribution function of a standard normal distribution.
13. The system of claim 11 , wherein the operations further comprise:
calculating a resulting expected membership advantage ρ α using sequential composition or Rényi differential privacy (RDP) composition; and
updating the at least one machine learning model using the calculated resulting expected membership advantage ρ α .
14. The system of claim 8 , wherein the calculating is based on a conditional probability of different possible datasets.
15. A non-transitory machine-readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to perform operations comprising:
receiving data specifying privacy parameters ε, δ which govern a differential privacy (DP) algorithm to be applied to a function to be evaluated over a dataset;
calculating, based on the received data, an expected membership advantage ρ α that corresponds to a likelihood of an adversary successfully identifying a member in the dataset, the calculating being based on an overlap of two probability distributions; and
applying, using the calculated expected membership advantage ρ α , the DP algorithm to a function over the dataset.
16. The non-transitory machine-readable medium of claim 15 , wherein the probability distributions are generated using a Gaussian mechanism with an (ε, δ) guarantee that perturbs a result of the function evaluated over the dataset, ensuring that membership advantage is ρ α on the dataset.
17. The non-transitory machine-readable medium of claim 15 , wherein the operations further comprise:
anonymously training at least one machine learning model using the dataset after application of the DP algorithm to the function over the dataset.
18. The non-transitory machine-readable medium of claim 17 , wherein the operations further comprise:
deploying the trained at least one machine learning model to classify further data input into the at least one machine learning model.
19. The non-transitory machine-readable medium of claim 15 , wherein the calculated expected membership advantage ρ α for a series of (ε, δ) anonymized function evaluations with multidimensional data is equal to:
CDF
(
1
2
2
ln
(
1.25
δ
)
ϵ
)
-
CDF
(
-
1
2
2
ln
(
1.25
δ
)
ϵ
)
wherein CDF is a cumulative distribution function of a standard normal distribution.
20. The non-transitory machine-readable medium of claim 18 , wherein the operations further comprise:
calculating a resulting expected membership advantage ρ α using sequential composition or Rényi differential privacy (RDP) composition; and
updating the at least one machine learning model using the calculated resulting expected membership advantage ρ α .Cited by (0)
No later patents cite this yet.
References (0)
No backward citations on record.